[juaristi22]

Summary

This PR takes the engine-free items of #1095 that María ruled on 29 October (Phase A), and the first re-bump past #1421 (Phase C): policyengine-uk 2.165.1. Main is merged in after #1420 and #1421, and the national and local target surfaces are regenerated on the pinned feed (Chronicle 53815b2). The spine and calibration items (Phase B) move to #1426, each measured in its own arm.

policyengine-uk 2.165.1

The lock moves from 2.164.2, which #1421 brought, to the newest release.

• Other investment income in market, HBAI, net and capital income (2.164.3, Count other investment income in market and HBAI income policyengine-uk#2736; 2.164.4, Count other investment income in capital income, and test it in every income measure policyengine-uk#2746). The build writes it from National Savings and gilt interest on FRS rows and from the SPI's other investment income on SPI rows. The engine already taxed it, and now also counts it in market income, HBAI income, household net income and capital_income. The stages that read those incomes:
  • The LCFS, NEED and NTS stages band recipients on the engine's HBAI income with its deductions added back, so their gross income now counts it, as the surveys' gross incomes do.
  • was_wealth, and was_lisa through its predictors, read household net income against WAS's DVTotinc_bhcR8, whose investment income covers every account and investment WAS asks about, National Savings bonds and gilts included. was_wealth also reads capital_income as a predictor, which now matches the donor's gross income from all financial investments (DVGIINVR8).
  • etb_services reads the engine's HBAI net income, and etb_vat its HBAI gross income less the ETB's direct taxes, against the ETB's disinc. The ETB's original income includes investment income, and its records come from the Household Finances Survey from 2018. Whether that survey's investment income counts National Savings and gilt interest is still to check against its documentation.
• Company car and car fuel benefits (2.165.0, Model the company car and car fuel benefits, with Class 1A NICs policyengine-uk#2750). The build writes none of the nine person inputs, so no one has a company car and both benefits and their Class 1A NICs are zero, as before. The inputs are signed at their default in UK_ENGINE_DEFAULT_INPUTS, for María's signature. The company car PR for the Autumn Budget dashboard supplies six of them.
• Tax-Free Childcare over a smaller UC award (2.165.1, Keep Tax-Free Childcare over a smaller UC award, and take no deductions from awards only a reform creates policyengine-uk#2754). A family that would claim UC keeps Tax-Free Childcare instead when the top-up is worth more than the award. No debt deductions come out of an award only a reform creates. Every new variable is a formula.
• No economic assumption moves. The vendored uprating pins and the bonus-sacrifice anchor change only their version stamps, and no gate digest moves.

Regeneration on feed 53815b2

The national and local references, membership and compile-parity receipts are regenerated on the merged tree, on main's feed pin (Chronicle 53815b2, #1420's re-pin). The vendored resources check against the full artifact unchanged. The local census, the local references, the local parity receipt and the coverage manifest regenerate unchanged.

• National surface: 1,409 active references, up from 1,403, and no row left without a fact (four before). The new rows are the three OBR council tax nation rows (D14e) and SLC's three nation rows (R4). The DfE rows, the R7 windows and obr.council_tax move as described below. SLC's three nation rows differ from the incumbent's devolved metrics only by their uprating to 2025.
• SPI uprating holds catch up with Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421: main's surfaces were generated before Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421 refreshed the HMRC uprating pins with the October EFO, so the regeneration moves 21 SPI holds. The eight self-employment amount band holds go from 1.1176 to 1.1115, as per-capita mixed income from 2023 to 2025 falls from 1.0384 to 1.0327. The 13 dividend holds by income band go from 1.088147 to 1.088151 (OBR per-capita GDP).
• Local surface: 27,168 active references, every status count unchanged. The signed deferrals fall from 94 to 77, as the support-floor deferrals now follow each target's nations.
• A window fix the regeneration found: R7's calendar-2025 window on dfi_ni.bus.passenger_receipts failed to compile, because its selector pins two services and the window counted series only from list-valued source keys. A dimension pinned to several values now names one series per value. The window of the sum is the sum of the windowed series, and a service short of a year, or with no facts at all, refuses the window. No other UK, US or NZ window reference pins several dimension values.

Targets

• OBR council tax by nation (D14e): the three nation rows are pinned to the UK and held as measured diagnostics.
• Declared geography wins (R8, R9): a selector's declared geography_id is returned before the inferred pin, and the authoring refuses a disagreeing pin.
• SLC devolved repayments (R4): slc.repayments.devolved_total becomes one row per nation. Wales binds. Scotland and Northern Ireland are held out by María's ruling, as their plan thresholds are not modelled. The three rows carry SLC's 2024-25 outturn to 2025 by policyengine-uk's OBR average earnings index, since repayments are a share of earnings above the plan thresholds (María's ruling of 30 October). This replaces the matched-period hold of 29 October, which declared the frame's own year: the solve honours a declared period only at the calibration year, or at the frame's year for a stored column, so it would have refused the Welsh row.
• Calendar-2025 window (R7): Scottish Child Payment, the TFC top-up, Northern Ireland bus receipts and the four SLC England repayment rows.
• Ledger observation_period: a reference can select its facts at a declared period while the target keeps its own.
• DfE (D5): the eight rows bind the January 2026 census at model year 2025. The three in_year_snapshot holds are retired.
• Window-ratio complement: a numerator can be its denominator less the listed cells.
• OBR council tax (D14f): binds the main-residence share: England's chargeable dwellings less second homes and empty homes, over chargeable dwellings (MHCLG, October 2025, about 0.968). Scotland and Wales take England's share, as a stated approximation.
• Signed exclusions: D3 keeps Scotland's three-plus-children rows out with a corrected reason. D18a signs the MSOA income deferral. D18c keeps Northern Ireland's UC area cells out.

Spine and stage records

• Carry manifest rewrites into the shared Stage/StageRecord producer graph #721: StageRecord carries rewrote, the stage's declared rewrites, in both the plan and the graph builds. The hand-written education_grants carve-out is gone, so Carry manifest rewrites into the shared Stage/StageRecord producer graph #721 can close when this lands.
• Support-floor deferrals follow each target's local-authority scope: the Welsh and Scottish families no longer defer the two English micro-authorities.

Gates

• Zero-weight strata (Adjudicate the UK spine grain at the ingest/calibration seam #688): both SPI-synthetic strata are capped at 0. The June caps of 100,000 counted the retired ten-clone grain, and the national seam ships no zero-weight row.
• Export surface (WS-E spine deferrals: known bugs and signed inconsistencies awaiting their increment #796): the all-zero incapacity_benefit_reported exclusion never matched the pinned reference, so it is dropped. The release boundary still drops the column.
• macro_realism (Swap the microcosm-built UK spine into the national build and retire the incumbent feed #757): not declared, and the battery's policy text records why. The solve binds the income components directly, and the national accounts' household disposable income differs in scope. The release contract refuses a not_applicable entry in this spec vintage.
• Take-up flags at their stages (María's ruling, "do both"): frs_take_up, frs_person_draws and frs_household_draws record their flags' shares at the stage's weights. Three assembled-phase gates hold them to their contract targets, at the terminal gate's bounds until a spine receipt measures them. The terminal uk_take_up_signal gate stays release-blocking.
• The battery, spine-part, release-cut and dense digests in microcosm-data's contract move with the manifest.

Disclosure floor of 3 (A1)

María ruled on 29 October to move every UK disclosure floor from ten source households to three, CD171's absolute floor (recorded on #1095). One constant, receipt_counts.UK_DISCLOSURE_FLOOR, now sets it for the stage receipts, the stage-health gates and the report tools. A withheld figure reads "<3", and counts from three to nine read "3-9", rounded record counts included (Vahid's review). The SPI tape's two SN 9422 conditions are unchanged. A guard test fails on any "<10" label left in the UK runtime, the UK spec, the gate notes or the tools. The statistical floors stay: the zero-support floor, the private-rent tool's ten-row group rule, and the SPI coherence tool's floor for medians and quantiles. The signed rules in local_small_cell_rule.json defer published cells, not survey cells, so they are unchanged.

Text

Stale tracker text in notes, docstrings and the vendored bus resources' purposes. The Incapacity Benefit leaf note now names the FRS 2024-25. The charter-H2 fixture and the release input coverage manifest follow the stage contract digests.

Not in this PR

• Promotion for the constant-id national line: pointer and certified-loader extension #823's default-load flip moved to its own PR, UK: make the promoted national line the default load (#823) #1432, by Vahid's review: it changes the default UK load for every consumer and depends on nothing here.
• 2.165.1's Tax-Free Childcare choice moves fewer than three source households on the 29 October spine, so it needs no data rule. The childcare refit (Phase B) runs on this lock.
• Ruled on 29 October: obr.sdlt keeps the households' share (the follow-up is in UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095); no Carer's Allowance stage gate until a caseload target binds; the HBAI check keeps its four statistics; SPI copies take their twin's contributory JSA and ESA only under the permitted-work limit (Phase B). Households' council tax on other property is Households' council tax on second homes and other property (engine and data) #1423.
• Still open: NS9c (explored separately) and the signed spine differences (after the childcare refit). D12b waits for the next calibration-year move, and certification against the promoted cut gets its own PR.
• Needs Chronicle first: Publish Chronicle package IDs in calibration diagnostics targets #661 (package ids in the consumer artifact).
• Overlaps: UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 and the company car PR change the same generated files (uk/gates.json and its digests, the release input coverage manifest, the charter-H2 fixture). The company car PR also locks policyengine-uk and re-pins the feed past chronicle#399. Whichever lands second regenerates them on the newer lock and pin. UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426's SPI pay-floor receipt still writes "<10", so the guard test makes whichever of the two lands second move it onto UK_DISCLOSURE_FLOOR.
• Phase B moves to UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 (each item measured): PR 2b, the stake level, HICBC opt-outs, the Carer's Allowance stage order, advanced ages, the benefit cap, and the rest of the approved plan.

Tests

• Engine tier on policyengine-uk 2.165.1: every UK engine test passes, run in full locally because CI stops at the first failure. Three needed updates: Tax-Free Childcare's top-up now lives in a shared helper, which still reads the spend the build routes; the company car's list price and capital contributions are person-level stocks that are no one's capital; and the graph build's stage records list each stage's rewrites beside their outputs (Carry manifest rewrites into the shared Stage/StageRecord producer graph #721). CI passed in full at 02df9df2b.
• Full engine-free suite at b46be0044 (673 files, in 30-file chunks): 22,969 tests, 233 skipped, counted by --junitxml. Sixteen fail:
  • ten fixed since: the SLC nation rows' declared period (four test_uk_measurement_period tests) and the pins and fixture rows that follow the regenerated surface (six). At 652835c1f, the 60 files that read the regenerated surfaces pass: 3,920 tests, 62 skipped;
  • six only on this machine: four Orrery tests (a missing Node package), test_engine_is_reported_unavailable_without_the_uk_extra (the uk extra is installed here) and test_spawned_workers_score_identically_to_sequential_cli (needs Python 3.14.7; this machine has 3.14.6).
• The "3-9" band (254be6d41): the tests of every rounded_count caller pass: 39 engine-free files (1,384 tests) and 8 engine-tier files (72 tests).
• After the Promotion for the constant-id national line: pointer and certified-loader extension #823 revert (1f46653b6): the 18 files that resolve or load a UK dataset pass, bar the live Hub pointer test, which needs private credentials (CI skips it). Main's version of that test expects the old release naming; UK: make the promoted national line the default load (#823) #1432 fixes it.
• Feed-reading tests on 53815b2 (the 15 feed-gated files, one process each, which CI skips without the feed): 235 tests, two skipped because the Scotland support and the authenticated national artifact are not on this machine. Two tests in test_uk_target_references.py still expected values from before D5 and R7; 02df9df2b moves them to DfE's January 2026 census and the calendar-2025 windows, each value checked against the feed's facts and the committed membership. The two drift checks that regenerate the vendored resources and the target surfaces from the feed were left out: the regeneration wrote those files, and the vendoring check passed in it.
• After Vahid's second review: f83d7a3c2 links the ruling's record from receipt_counts and the changelog, and its tests pass. ddd82aa2f adds a test that a pinned service with no facts refuses the window; the 460 tests of test_ledger_targets.py pass.
• Checks: tools/ci_test_plan.py verify, tools/spec_engine_coverage.py --check, the vendoring check against the full artifact, and ruff on the edited files.

🤖 Generated with Claude Code

[vahid-ahmadi]

@juaristi22 this is an early look at 5a944e449, not a full review, since the body says it isn't ready. It covers must and should items only, and nothing on your own pending list. I'll do a full pass after main merges in and the references regenerate.

1. must: an engine-uk failure the body doesn't list. tests/engine/uk/test_uk_graph.py:84 still asserts set(nonzero_share) == set(stage.outputs). But _graph_stage_records now adds stage.rewrites to shares (Carry manifest rewrites into the shared Stage/StageRecord producer graph #721). So age_tail fails in engine-uk on both Pythons (run 40917095981, at this head). Regenerating won't fix it. The assertion needs set(stage.outputs) | set(stage.rewrites). The body's engine-tier line covers only the charter-H2 parity test, so this one was missed.
2. should: rounded_count now prints exact counts from three to nine (receipt_counts.py:103). The module docstring still says "no exact count that ... narrows a small group by difference". extreme_count and support_band print "3-9" there instead. Printing exact counts just above the floor is where differencing bites. That matters more now that the floor is three rather than ten, CD171's working minimum against secondary disclosure across outputs, as the docstring says. I'd keep "3-9" in rounded_count too. I'm not reopening the ruling. Please link the UK: every open data item: uk-data parity, engine locks, the local surface, consumption and high-value homes #1095 comment that records it from the module and the changelog, and say there that the SN 9422 conditions (HMRC sees SPI-based material ten working days ahead) are unchanged.
3. should: UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 adds new "<10" sites. The SPI pay-floor receipt in UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 hardcodes ten in several places. Whichever of the two lands second should move them onto UK_DISCLOSURE_FLOOR. Otherwise one receipt stays on the old floor without anyone noticing. I've raised it on UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426 too.
4. should: ship the Promotion for the constant-id national line: pointer and certified-loader extension #823 default-load flip on its own. Pointing load("uk") at latest-national.json and retiring variant="national" affects every consumer. It doesn't depend on Budget 2026 slower data fixes, PR C of #1418: electricity at the ONS level, pension income by area #1420, Budget 2026 ISA data, PR B of #1418: ISA subscriptions per adult, the Lifetime ISA anchor, additive ISA income, policyengine-uk 2.164.2 #1421 or the regeneration. As its own small PR it could land now, and it wouldn't be held behind this one. The old 823-national-variant-until-promoted.changed.md entry, which describes the opposite state, should go in the same change.

The D14f main-residence share (0.968) is covered on #1423: #1423 (comment).

What I ran. I read the diff at 5a944e449 against its base fddbc7073 and the failing CI logs. I didn't run anything locally. I didn't review the target, gate or ledger changes in detail, and I'll leave those for the full pass.

🤖 Generated with Claude Code

[juaristi22]

Thanks for the independent review of 0a69ef4b5. Every finding is addressed at c0dd6eae2, except finding 4, which belongs to whichever of #1422 and #1430 lands second.

Fix before merge

1. Regional home values and "3-9": the gate reads any banded count at its bounds. A region of three to nine rows counts as three, the stage's own banded row count is read at its top, and only a region printed "<3" is flagged, as before. The fixture's counts are now ones the receipt can print, and a new test covers "3-9" (7a95f37a7).
2. The pay-ratio clamp: it stays at ten, now PAY_CONCEPT_RATIO_MINIMUM_SUPPORT, as a statistical floor: on fewer households a ratio past its bounds is noise, and stopping the build on it is what Vahid's second review of Build the national + local-area pair on policyengine-uk 2.151.0, and drop UK model-figure withholding #1396 ruled out. The gate now reads a ratio whose support band is under ten as unmeasured, so a held ratio is flagged rather than judged (7a95f37a7).

Periods and windows

3. observation_period across restamps: restamp_reference_period moves a declared observation period by the same number of years as the target period. The calibration compile, the local compile and the childcare fit tool restamp through it. So the DfE rows read the January census after their model year at any period, and --target-period and --vintage-override select the matching census. The test now restamps the row instead of nulling the field (42913ab5e).
4. UK: company cars from the FRS job table (dashboard #316) #1430's feed: right; this falls to the second PR to land. The UK: company cars from the FRS job table (dashboard #316) #1430 session has the three nation rows to add to source_table_by_opening_year when it merges main.
5. The engine-defaults guard: it also reads every stage's declared outputs and rewrites from source_stages.json. UK: company cars from the FRS job table (dashboard #316) #1430's company_cars stage will fail it until its six written inputs leave UK_ENGINE_DEFAULT_INPUTS (7a95f37a7).
6. Series per pinned value: a window over several pinned members now needs exactly one complete series per member. Your case of two ulsterbus series and no metro_glider refuses, with a test (42913ab5e).
7. The nation diagnostics' basis: the three nation rows now take obr.council_tax's main-residence share, so they sum to the bound row and each reads its nation on the same basis. The notes, exclusion reasons and changelog say so (c0dd6eae2).

Gates and tests

8. Withheld cells in the take-up stage gates: they are flagged unchecked, never failed, as the sibling gates do under María's ruling of 26 October. The terminal gate still measures every flag at the calibrated weights, so nothing goes unmeasured (7a95f37a7).
9. childcare_copy_take_up: it now records the five flags it re-draws, and a fourth assembled gate holds them to their targets. The changelog says so (7a95f37a7).
10. Test gaps: frs_take_up's receipt is read in its stage test. A new test runs each take-up stage gate through the battery's stage_health binding, as the spine battery does. The copy stage's gate is tested, and so is "3-9" in the home values gate (7a95f37a7).
11. The floor guard: it now scans tools/** and microcosm-data too. A second guard fails on any comparison, constant, default or prose that keeps a floor of ten outside the statistical floors it names (7a95f37a7).

Texts

12. Floors of ten: moved to three are the high-value-homes support gate and its default, and the three tools' SDC defaults. Kept as statistical floors, each now saying so, are the pay ratio's ten, the SPI transport's ten rows, the regional renter minimum, the SPI coherence tool's cells and the private-rent tool's groups. The stale prose is restated.
13. Replaced designs: restated are the R4 changelog, the parity tool's DfE and council tax rationales, the DfE contract note, the local census's MSOA entry (now D18a's reason) and the generator's SLC comment (c0dd6eae2). The regenerated receipts carry them.
14. The complement's metadata: the formula now reads numerator / denominator, matching the recorded numerator, and the test checks that the recorded pieces give the value back (42913ab5e).

The national and local surfaces, both parity receipts and the local census are regenerated on Chronicle 53815b2 in a follow-up commit. Until it lands, CI's checks of the generated surfaces will fail.

[juaristi22]

Re-review of #1422 at 9d25e4579

This re-reviews the fixes in c0dd6eae2, and the regeneration in 43c3da74e and 9d25e4579, against my review at 0a69ef4b5. One reviewer re-ran each earlier finding's reproduction, a second read the fix commits for new defects, and I reproduced the blocker below. Paths are relative to packages/microcosm-build/src/microcosm/build/ unless they start with packages/, tools/ or changelog.d/.

Verdict: one new blocker, from the item 9 fix. The childcare copy gate is declared at a phase that runs before its stage, so it stops every licensed spine build (1). Item 4 is deferred as agreed, and the other round-1 items are fixed, with small residuals.

Fix before merge

1. The childcare copy gate runs before its stage.

• uk/gates.json:1002 puts uk_stage_childcare_copy_take_up_signal in the assembled phase, and packages/microcosm-data/src/microcosm/data/contract.py:567 does the same. UK_SPINE_GATE_SCOPE arms it in the spine battery.
• The assembled boundary is frs_brma, stage 13 of 50, but childcare_copy_take_up is stage 48. Each boundary offers evidence only for the stages that have run (uk_runtime/spine_build.py:1326, and uk_runtime/graph_evidence.py:251 in graph mode).
• So the binding raises "stage_evidence has no receipt object for 'childcare_copy_take_up'" (uk_runtime/battery_bindings.py:327), and the gate fails closed. Because it is release-blocking, enforce stops the build at the boundary, on dev and release-candidate builds alike.
• A synthetic smoke build records the failure and carries on, because the entry sets population_fact_check. So integration-uk passes.
• Why the tests miss it:
  • the binding test skips the copy stage (packages/microcosm-build/tests/engine_free/uk/test_uk_take_up_gate.py:449);
  • the copy stage's own test checks only that no flag is omitted or undeclared (test_uk_childcare_copy_take_up.py:383). On that test's frame the gate would fail anyway ("max_free_entitlement_hours_used is constant"), since the frame's age-1 usage rate is 1.
• No other assembled-phase stage gate reads a stage after the boundary.
• Fix:
  • Move the gate to transferred, as the student_loans and isa_subscriptions stage gates are.
  • Carry the change through contract.py:567, packages/microcosm-data/tests/engine_free/shared/test_contract.py:387, the digests and the changelog's "four assembled-phase" gates.
  • Add a test that every assembled-phase stage_health gate reads a stage at or before UK_SPINE_ASSEMBLED_FINAL_STAGE, so this can't come back.

Round-1 items

#	Status
1. Home values gate on "3-9"	Fixed. _count_bounds (uk_runtime/stage_health.py:4641) reads each region's band at its bottom. 2, 3, 9 and 10 rows all pass, and a new test covers "3-9".
2. Pay-ratio clamp	Fixed. A true ratio of 1.5 on 5 or 9 households is now flagged pay_concept_ratio_measured: False. Residual: the receipt still prints the held 1.1 as pay_concept_ratio (uk_runtime/frs_annual_income.py:943-945), and the gate repeats it beside the flag. Only the support band shows that it was held.
3. observation_period and restamps	Fixed at the three named sites, which now call restamp_reference_period. tools/diagnose_uk_uc_matrix.py:389 still restamps the old way on its --references route. See also 2 and 3 below.
4. Nation rows on #1430's feed	Deferred to whichever PR lands second, as your reply says. The rows are now scaled windows, still without a source_table_by_opening_year map. A trial merge with #1430 conflicts only in three notes hunks and five digest hunks. The rows read the same MHCLG facts that obr.council_tax already needs.
5. Engine-default guard	Fixed. Run on #1430's stages and fixture, it flags the six company_cars outputs. So #1430 will fail it until they leave the defaults list, which is the point.
6. One series per pinned value	Fixed and tested.
7. Nation basis	Fixed. The three rows carry obr.council_tax's main-residence ratio, and 43c3da74e regenerates them at 0.96767 × the window. Scotland and Wales take England's share, which the notes state.
8. Withheld cells	Fixed and tested.
9. Copy stage gate	Added, but at the wrong phase: see 1 above.
10. Test gaps	Fixed: the three named gaps are filled. The copy gate still has no end-to-end test (1).
11. Floor guard	Partly fixed. It now scans tools/** and microcosm-data, and it catches < 10 on labelled lines. Still missed: unlabelled forms such as distinct_count(…) < 10, np.unique(x).size < 10, >= 10, minimum=10, "< 10" with a space, and "at least ten". On #1426's spi_income.py with only the labels fixed, it lets :961 and :1790 through. Allow-list: it matches by substring, so the uk/spec/sources.yaml:1717 needle also exempts any line containing it. Scope: of the uk/*.json files it reads only gates.json (see 4 below).
12. Floors of ten	Moved, or labelled as statistical floors. Two texts remain: uk_runtime/spi_rank_transport.py:622-626 still withholds positive_rows under its 10-row constant, citing CD171 5.2.1, and uk/spec/sources.yaml:1717 still says a positive-row count under 10 is withheld. Both contradict the constant's new label. They are stricter, so nothing leaks.
13. Stale texts	Fixed, the receipts included after 43c3da74e. One more of the same kind: tools/build_uk_ledger_compile_parity_signed_differences.py:617-622 still calls the ONS small-area income rows a "Signed deferral", pending MSOA-to-LA aggregation. The regenerated local parity receipt repeats that on 903 rows, while the local census now keeps them out as circular (D18a).
14. Complement metadata	Fixed. The test checks the numerator, ratio and formula, but not the denominator.

New in the fix commits

2. restamp_reference_period shifts only a leading four-digit year (ledger_targets.py:4667).

• "2024-25" becomes "2025-25", which the reference then refuses, and "FY2024" raises even at a zero shift.
• The national compile calls it outside its try block (uk_runtime/ledger_targets.py:494 and :1430). So one such row aborts the compile instead of being recorded as unsupported.
• No committed row has these forms, so this is latent.

3. The authoring route doesn't restamp.

• target_reference_authoring.py:1177 and :1415 set period from --period, but :1202-1204 and :1429-1431 copy observation_period unchanged.
• So authoring the DfE rows at --period 2026 gives "2026-01", where the committed rows restamped to 2026 read "2027-01".
• With tools/diagnose_uk_uc_matrix.py:389 (item 3 above), that makes two routes that bypass the helper.

4. Old floor-of-ten texts in uk/*.json files the guard doesn't read. Four of them still justify a choice by the floor of ten:

• uk/calibration_measure_exclusions.json:79 says twice that rows under 10 source households are withheld;
• uk/target_reference_signed_exclusions.json:168 and :181, and uk/local_small_cell_rule.json:44, sign out the North East and Wales £2m rows partly because a public sign-off "would narrow an estimate expected to rest on fewer than 10 records".

Restate them, or list them as kept.

5. The home values gate reads rounded counts as exact (low; uk_runtime/stage_health.py:4641).

• _count_bounds treats a count of 10 or more as exact, but the producer rounds it to tens. Ten regions of 15 rows print 20 each, and against 150 stage rows that fails as "more rows rewritten than the stage carries".
• It needs nearly every row to be an England-and-Wales owner, so only synthetic frames hit it.

Checked and found true

• Contract: the digests match the entries, and the gate battery and contract tests pass.
• One-series check: it covers the four multi-value windows, and none collide.
• The regenerated surfaces move only as described:
  • the three nation rows;
  • the DfE and obr.council_tax reasons in the parity receipt;
  • the ONS small-area income entry in the local census.
• CI on 9d25e4579: lint, wheels and integration-uk pass. The engine jobs were still running.

Disclosure-control items went to the author separately.

🤖 Generated with Claude Code

[juaristi22]

Thanks for the re-review of 9d25e4579. Everything below is at 5b693b736.

Fix before merge

1. The childcare copy gate's phase: it now runs at the transferred phase, as the WAS, LCFS and ETB stage gates do, since its stage runs after the assembled boundary (frs_brma). The contract's mirrors, the changelog and the digests follow. A new test fails on any assembled-phase stage gate that reads a stage after the boundary. The copy gate now passes end to end through the battery's stage_health binding, on a frame with children under 1, so the under-2 usage varies.

Round-1 residuals

• 2: the gate's details now say pay_concept_ratio_held_to_bounds when the stage held the ratio, beside pay_concept_ratio_measured: False. The receipt still prints the held value, with its support band.
• 3: tools/diagnose_uk_uc_matrix.py restamps through restamp_reference_period.
• 11: the guard now catches a comparison with ten in any form (< 10, >= 10, 10 <), a keyword or default of ten, "< 10" with a space, and "at least ten". It reads the UK tools, microcosm-data and the hand-written UK JSON. Each ten it keeps is allowed only by the phrase that contains that match, not by its line. A probe with your unlabelled forms fails it.
• 12: the SPI transport now rounds its positive-row count as every receipt count is ("<3" for one or two rows), and the spec says the ten rows are a statistical floor for the distances.
• 13: the parity tool now gives D18a's reason for the ONS small-area income rows. The local parity receipt picks it up at its next regeneration, which runs when the machine is free.
• 14: the complement test checks the denominator too.

New in the fix commits

• 2: restamp_reference_period moves both ends of an annual range ("2024-25" to "2025-26") and a prefixed year ("FY2024"), and leaves any spelling alone at a zero shift. A row it can't move is recorded as unsupported instead of aborting the compile.
• 3: the authoring tools move a contract's observation period with the authored period. The UK generator declares the contract's base, 2025, so authoring at --period 2026 gives "2027-01", as the restamp does.
• 4: the U3 adjudication now states the floor of three beside its history. The two signed £2m exclusions and the local small-cell rule are kept as they are: each estimate rests on about one household weight, so the reason holds at the floor of three too, and their texts are copied into 603 generated membership rows. The guard names them as kept.
• 5: the home values gate now reads a rounded count as the range of counts that round to it, so ten regions printed 20 pass against 150 rows. A test covers it.

The touched test files pass locally, and CI runs the full engine-free and engine tiers.

[juaristi22]

Third review of #1422 at 5b693b736

This re-reviews 5b693b736 against my second review at 9d25e4579. As before:

• one reviewer re-ran each earlier finding;
• a second read the commit for new defects;
• I checked the main points myself.

Paths are relative to packages/microcosm-build/src/microcosm/build/ unless they start with packages/, tools/ or docs/.

Verdict: the blocker is fixed, and nothing here blocks a build. Before merge:

• regenerate the local parity receipt, since CI can't see that it's stale (item 13 in the table);
• reconsider the new try/except in the compile; my last review pointed the wrong way there (1 below);
• tighten the rounded-count bounds, which the commit loosened along with a test (2 below).

The rest are latent edge cases and texts.

Round-2 items

#	Status
1. Copy gate phase	Fixed. uk/gates.json:1002 is transferred, and the mirrors, digests and changelog follow. The new test (test_uk_stage_health.py:3387) would have failed at 9d25e4579. Through the battery, the copy gate now resolves at the transferred boundary and passes on the infant frame. One gap: live spine gates run in graph mode, which hard-codes "frs_brma" (uk_runtime/graph_evidence.py:243-245). The test reads UK_SPINE_ASSEMBLED_FINAL_STAGE, whose only user, _run_plan_with_spine_sampling (uk_runtime/spine_build.py:1286), has no callers. So the test guards the live boundary only while the two agree. Having graph_evidence.py read the constant would tie them together.
2. Restamp helper	Fixed for the named forms: "2024-25", "FY2024" and "2026-01" all move correctly. Latent edges are in 3 below.
3. Authoring route	Fixed for the national generator. With base 2025, the 8 DfE rows give "2025-01", "2026-01" and "2027-01" at 2024, 2025 and 2026. The local generator declares no base, which is harmless while no local row has an observation period.
4. Floor-of-ten texts	Partly fixed. :79 is restated, and the kept texts are named. Still wrong: the support-floor reason at uk/calibration_measure_exclusions.json:76 still gives the floor as ten ("fewer than 10 distinct source households"), and the guard misses it (4 below).
5. Rounded counts	Fixed for the round-2 case, but loosened elsewhere (2 below).
Item 2 residual: pay ratio	Fixed. But: the new pay_concept_ratio_held_to_bounds is set from the support band alone (uk_runtime/stage_health.py:1317-1327), and the stage clamps only a ratio outside the bounds. So an in-bounds ratio the stage never moved is also marked held: 1.02 on 5 households, the stage test's own case (test_uk_frs_annual_income.py:544).
Item 3 residual: diagnose tool	Fixed. No other UK restamp bypasses the helper.
Item 11: floor guard	Partly fixed. A labels-only fix to #1426's spi_income.py now fails it, which is the case that mattered. The remaining gaps are in 4 below.
Item 13: parity reason	Tool fixed, receipt stale. uk/ledger_compile_parity_local_incumbent_2025_signed_differences.json keeps the old "Signed deferral" reason on 903 rows. No CI test compares reasons: the local receipt replay matches names, kinds and values, and the only regeneration test is national and skips without the feed. So it merges stale unless it is regenerated first.
Item 14: complement	Fixed.

New in 5b693b736

1. The compile's new try/except turns a contract error into an unsupported row (uk_runtime/ledger_targets.py:494-500, :1436-1442).

• The same function says at :507-510 that a contract error should "refuse the whole compile rather than report it as an unsupported row". An observation period that can't be moved is one.
• At the calibration year, full_targets.py still refuses, so a build stops. At validation or parity periods, though, the row becomes ledger_absent, and a regeneration would sign it with the generic reason.
• My round-2 point 2 pointed the wrong way here. The abort was the right outcome; it only needed to name the row. Raising with the reference's name, as :507-510 does, fits the codebase's rule.
• No test covers the branch.

2. The rounded-count bounds are too wide at 10 and 1,000, and a test was changed to fit (uk_runtime/stage_health.py:4659-4662).

• rounded_count prints 10 only for counts of 10-14, since 5-9 print "3-9". It prints 1,000 for 995-1,049. But _count_bounds returns (5, 14) and (950, 1,049).
• So impossible receipts now pass. For example, ten regions printed 10 (at least 100 rows) against a stage printed 90 (at most 94). That failed at 9d25e4579.
• test_uk_stage_health.py:2802 moved its mutation from rows=90 to rows=40 to keep failing. The comment and assertion at :2844-2849 say nine regions printed 10 plus one "3-9" pass against rows=50, though they need at least 93 rows.
• Fix: use max(v - 5, 10) below 1,000 and max(v - 50, 995) from 1,000. Only synthetic frames reach these values.

3. _shift_period_years disagrees with the ledger's own period parser (ledger_targets.py:4711-4718; latent).

• Month or range? For 2000-2011, YYYY-MM with MM equal to the next year's two digits is read as an annual range. So "2011-12" (December) moves to "2012-13", and "2005-06" moves to "2006-07".
• Refused spellings: the ledger accepts "2024-2025", "2024_2025" and "2026-1", but the helper refuses them at any nonzero shift.
• Unparseable target period: "2025Q1" still turns "2026-01" into "1-01".
• Suggestion: classify the spelling with _normalize_period_value (ledger_targets.py:5119). All committed rows are "2026-01" at 2025, so nothing is affected now.

4. The floor guard still misses common forms and has generic allow-list needles (packages/microcosm-build/tests/engine_free/uk/test_uk_receipt_counts.py:173-183).

• Prose: in digits, prose matches only when one of seven nouns follows "10". So it misses:
  • the live text at :76 ("10 distinct source households");
  • "Fewer than 10 FRS source households" (uk/uk_population_targets.json:682, :35132);
  • "fewer than 10 awards" (uk_runtime/frs_spine.py:2016).
• Code: it misses keywords not named minimum* (min_rows=10, min_count=10, threshold=10), > 9 and <= 9, and YAML min_rows: 10.
• False positives: it flags < 10_000, < 10% and format specs such as f"{x:<10}".
• Allow-list: the needles are generic, so a new if value < 10: in receipt_counts.py or stage_health.py passes.
• Dead needle: "about one household weight" (:142) contains no ten, so it can never match.

So the reply's "a comparison with ten in any form" is overstated. Either tie each kept entry to its file and exact phrase and widen the prose match, or say in the test which forms it checks.

5. Small items:

• No changelog fragment describes restamping observation periods or the authoring base.
• The base 2025 is hand-set in tools/generate_uk_target_references.py:147-149, and no test ties it to the committed references' period.

Checked and found true

• Phase move: every boundary (driver and graph) offers the copy stage's receipt at the transferred phase. Certification checks each entry's phase, and the spine scope and all mirrors include the gate. No other assembled-phase gate reads a later stage, in this tree or UK: company cars from the FRS job table (dashboard #316) #1430's. The old digest appears nowhere.
• Coverage manifest: its 23 changed lines are the new source_stages.json hash, and tools/build_uk_release_input_coverage_manifest.py --check reports it current.
• Compile: an unsupported reference at the calibration year still makes full_targets.py refuse, so no bound row is dropped silently.
• Tests: the touched test files all pass, along with the contract and digest pins.
• CI on 5b693b736: lint, wheels and integration-uk pass. The engine jobs were still running.

Disclosure-control items went to the author separately.

🤖 Generated with Claude Code

[juaristi22]

Thanks for the third review, of 5b693b736. Everything below is at 821e342d1.

Before merge

• The local parity receipt (item 13): regenerated in ce9344c68. Its 903 ONS small-area income rows now give D18a's reason, and nothing else moves. bd9839bf2 also regenerates the H2 fixture, which 5b693b736 had made stale.
• 1. Restamp errors: reverted to stopping the compile. A reference whose observation period can't be moved now raises, naming the row, as the function's other contract errors do. A test covers the branch.
• 2. Rounded-count bounds: a printed count now stands only for counts that print as it: 10 for 10 to 14, and 1,000 for 995 to 1,049. The home values test is back to rows=90 for its mutation. The "3-9" case now says that nine regions printed 10 and one of three need 93 rows, so it fails against a stage printed 80 and passes against one printed 90.

Round-2 items

• 1, the graph boundary: UK_SPINE_ASSEMBLED_FINAL_STAGE now lives in graph_evidence.py, and the graph's spine gates read it. So the stage-gate test checks the boundary the build runs.
• 4: the support floor's reason now gives the floor of three.
• Pay ratio: the receipt says whether the stage moved the ratio onto a bound (pay_concept_ratio_held, withheld beside a withheld ratio), and the gate reports pay_concept_ratio_held_to_bounds from it. So 1.02 on five households is unjudged, but not reported as held.

New in 5b693b736

• 3. Periods: _shift_period_years now classifies a spelling as _normalize_period_value reads it:

  • "2011-12" and "2005-06" are months, so only their year moves;
  • "2024-2025" and "2024_2025" are ranges, so both ends move;
  • "2026-1" moves its year;
  • a target period without a year, such as "2025Q1", refuses.

• 4. The floor guard: each kept ten is now named by its exact phrase and the number of matches it covers in its file. A new if value < 10: fails, and so does an entry that matches nothing, so the dead one is gone. The guard now finds:

  • "fewer than", "under", "below", "less than" or "at least" before "ten" or "10", whatever noun follows ("10 distinct source households", "10 awards");
  • keywords, keys and defaults named min*, minimum* or threshold* set to ten, in Python, YAML and JSON;
  • > 9 and <= 9.

  It skips format specs, 10_000, 10% and 10 * x. Its docstring lists the forms it checks, rather than claiming every form. The two "Fewer than 10 FRS source households" notes and "fewer than 10 awards" are named as facts that hold at the floor of three.

• 5: UK_TARGET_REFERENCE_PERIOD (2025) names both the period the two generators author at and the base of the contract's observation periods. A test ties it to the committed references, and the observation-period changelog now describes the restamp and the authoring base.

The touched test files, the H2 parity test and the H2 fixture check pass locally. CI runs the full tiers.

[juaristi22]

Fourth review of #1422 at 821e342d1

This re-reviews 8f7d69204, ce9344c68, bd9839bf2 and 821e342d1 against my third review at 5b693b736. As before:

• one reviewer re-ran each earlier finding;
• a second read the commits for new defects;
• I checked the main points myself.

Paths are relative to packages/microcosm-build/src/microcosm/build/ unless they start with packages/ or tools/.

Verdict: the round-3 items are fixed, and so are the two CI failures since 5b693b736. Nothing here blocks a build. Two of the fixes regress latent cases, and in both the author followed my round-3 suggestion: periods under a typed selector (1) and the floor guard on a sentence-final "ten." (2). Item 3 is a trap for whoever next moves the reference year.

Correction to my last review: it said the engine jobs at 5b693b736 were still running. They then failed:

• engine-free on the payload comparator's CLI test, stale since 7a95f37a7 and also failing at 9d25e4579;
• engine-uk on the H2 spine parity test.

8f7d69204 and bd9839bf2 fix them. engine-free runs with --maxfail=1, so each run shows only its first failure.

Round-3 items

#	Status
Local parity receipt	Fixed. Only the 903 reasons changed, and they now equal the tool's text. Can recur: nothing in CI compares the local receipt's reasons with the tool.
H2 fixture	Fixed. One stage-contract hash and the transport text changed, and every node's hash recomputes.
1. Contract errors	Fixed. Both compiles raise and name the row, and the national branch's test fails if that is reverted. Gaps: the local branch has no test, and its message says "UK target reference" where its sibling says "UK local reference".
2. Count bounds	Fixed. I checked to 20,000: the bounds now equal exactly the counts each printed value stands for, and rows=90 is back. Gap: nothing pins the 1,000 edge; reverting it to v - 50 passes every test.
3. Period spellings	Partly: a regression under typed selectors (1 below).
4. Floor guard	Partly: a regression on sentence-final tens (2 below).
5. Base and changelog	Partly (3 below).
Graph boundary	Fixed. The graph's spine gates read the constant from uk_runtime/graph_evidence.py. Residuals: uk_runtime/graph.py:1719-1720 still names "frs_brma" for the checkpoint node, though a divergence fails loudly. The test imports the constant through spine_build's re-export (test_uk_stage_health.py:3394), which breaks if the dead plan path is removed.
Support floor reason	:76 now gives the floor as three.
Pay ratio	Fixed. 1.5 on five households reads as held, 1.02 on five doesn't, and ten or more behave as before. A receipt from before this commit just lacks the flag.

New in these commits

1. Periods shift wrongly under a typed selector (ledger_targets.py:4716, against :5063-5068).

• The helper now classifies a spelling without a type. The exact-match path, however, reads an observation period with the selector's period_type.
• Under a fiscal_year selector, "2000-01" to "2011-12" are fiscal ranges. So "2011-12" +1 now gives "2012-12", which no fiscal-year fact matches; round 3 gave "2012-13".
• My round-3 suggestion to classify with _normalize_period_value caused this. The helper should classify with the reference's period_type when it has one.
• A target period of "FY25" still turns "2026-01" into "26-01".
• It's latent: every committed observation period is "2026-01", with no typed selector.

2. The floor guard's new lookahead misses a sentence-final ten (packages/microcosm-build/tests/engine_free/uk/test_uk_receipt_counts.py:213).

• (?![\d_.%]) rejects a following full stop. So these now pass, though round 3 caught them:
  • "fewer than ten." or "under 10." at the end of a sentence;
  • constants and keys set to 10.0.
• Fix: (?![\d_%]|\.\d) would still skip "10.5" without losing them.
• Wrapped prose: it still misses prose split across a line. tools/uk_private_rent_receipt.py:416-417 still says "fewer than ten" across a line break, so the guard can't see it.
• False positives: it still flags f"{count:>9}" and minutes=10. These fail loudly, so they're harmless.

3. One constant plays two roles (uk_runtime/ledger_targets.py:480-484).

• UK_TARGET_REFERENCE_PERIOD is both the generators' default --period and the contract's observation base.
• The trap:
  • Moving the references to 2026 the right way, with --period 2026, gives "2027-01" but fails the new test (test_uk_target_references.py:2239).
  • Bumping the constant instead passes the test, but it moves the base too. The DfE rows then read "2026-01" at 2026: the model year's own census, the error D5 fixed.
• Suggestion: keep the base separate, since it belongs to the contract, and check that each observation period is the January after it.
• The test also checks only the national references. The changelog's "classified as the ledger's own period parser reads its spelling" doesn't hold for typed selectors (1).

Checked and found true

• Regenerated files: the receipt and fixture diffs contain only the stated changes. The touched modules import with no cycle, and the graph's stage order matches the declared roster.
• Restamps: the try wraps only the restamp, so fact gaps are still recorded as unsupported.
• Manifest: the coverage manifest is current.
• Tests: the touched test files pass.
• CI on 821e342d1: every job passes.

Disclosure-control items went to the author separately.

🤖 Generated with Claude Code

[juaristi22]

Thanks for the fourth review, of 821e342d1. Everything below is at 0da33d0cd. Noted on the engine jobs: 8f7d69204 and bd9839bf2 fixed both failures.

New in the round-3 commits

• 1. Periods under a typed selector: the restamp and the authoring tools now read an observation period under the selector's period_type when it declares one, as the exact-match path does.

  • Under fiscal_year, "2011-12" moves to "2012-13". Under month, or with no type, it moves to "2012-12".
  • "2026-01" spells no fiscal year, so a fiscal-year selector refuses to move it.
  • A target period has to give a four-digit year, so "FY25" refuses rather than moving "2026-01" to "26-01".
  • Tests cover each case.

• 2. The floor guard:

  • A ten is now 10, 10. at the end of a sentence, or 10.0. The guard still skips 10.5, 10% and 10_000.
  • Prose matches across a line break, a comment marker or a string's quotes. That found three wrapped tens:
    • the rent receipt's "fewer than ten". It was stale: the tool already withholds below the floor of three (MIN_SOURCE_HOUSEHOLDS), and its docstring now says so;
    • the pay ratio's "one to nine households", which is a statistical floor;
    • the SPI tape's "at least ten working days" for HMRC's review, which is a licence term.
  • The last two are kept by their exact phrases.
  • {count:>9} and minutes=10 no longer match: the pattern for constants is case-sensitive now and skips MINUTE.

• 3. The base: UK_OBSERVATION_PERIOD_BASE (2025) now holds the contract's base, separate from UK_TARGET_REFERENCE_PERIOD, the generators' default --period. Two tests replace the old one:

  • the contract's observation periods are each the January after the base;
  • for the national and the local references, each reference whose contract target declares an observation period reads the January after its own period, and no other reference declares one.

  So authoring at 2026 gives "2027-01", which passes once the default moves with it, and moving the default leaves the base alone. The changelog now says the period is read under the selector's type.

Round-3 gaps

• Local parity receipt: a new test replays the tool's reason notes over both committed parity receipts. It fails on any reason the tool would now write differently, and changing one of the tool's texts marks its rows stale. The national replay needs the feed, so CI didn't check either receipt's reasons before.
• Contract errors: the local compile's message now says "UK local reference". A test covers that branch with a real fiscal-year reference that can't move.
• Count bounds: a test checks every count up to 20,000. Each printed count's bounds equal the lowest and highest counts that print as it, and putting the 1,000 edge back to v - 50 fails it.
• Graph boundary: UK_SPINE_ASSEMBLED_FINAL_STAGE now lives in uk_runtime/graph.py, which builds the checkpoint node from it. graph_evidence.py, spine_build.py and the stage-gate test import it from there. Each touched module imports on its own in a fresh interpreter.
• Pay ratio: a receipt from before 821e342d1 lacks the flag, so the gate doesn't report its ratio as held. A thin ratio still reads as unmeasured. No change.

The touched test files pass locally, and so do the H2 fixture check, the H2 parity test, the coverage manifest check and the gate digests. CI runs the full tiers.

[juaristi22]

Fifth review of #1422 at 0da33d0cd

This re-reviews 0da33d0cd against my fourth review at 821e342d1. As before:

• one reviewer re-ran each earlier finding;
• a second read the commit for new defects.

Paths are relative to packages/microcosm-build/src/microcosm/build/ unless they start with packages/ or tools/.

Verdict: the round-4 items are fixed, CI is green, and nothing blocks a build. Two of the fixes regress latent cases:

1. under the default match policy, typed selectors now refuse to restamp. My round-4 suggestion pointed there, again;
2. the floor guard's case-sensitive constant pattern drops lowercase floor names.

Round-4 items

#	Status
1. Typed periods	Fixed for exact. A fiscal "2011-12" now moves to "2012-13", and "FY25" refuses. But see 1 below.
2. Floor guard	Fixed. Sentence-final tens, 10.0 and wrapped prose are caught, and the three wrapped tens are real and counted. But see 2 below.
3. Observation base	Fixed. Every mutation I tried fails the new tests, including the old trap with both constants at 2026.
Parity-reason replay	Partly. It runs feed-free in CI and catches changed tool texts and stale ruled reasons. Not caught: rows that keep the classifier's generic text (318 national, 288 local), and a removed national rule. So "fails on any reason the tool would now write differently" overstates it.
Local message	Fixed, but its test pins 1's regression (below).
Count bounds	Fixed. Every edge mutation fails the test.
Graph boundary	Fixed. There is one definition, in uk_runtime/graph.py, with no import cycle, and #1430's head merges with it cleanly.

New in 0da33d0cd

1. Typed restamps now refuse what the default policy reads (ledger_targets.py:4691-4695; target_reference_authoring.py:151).

• Mechanism:
  • The restamp reads the observation period under the selector's type for every policy. Only exact reads it that way.
  • The default latest_not_after, which every committed DfE row uses, reads it untyped as a cutoff (:4766, :4853-4854, :4890).
• Example: an academic_year or fiscal_year selector with observation "2026-01" and the default policy selects AY2025-26 at 2025.
  • Round 4 moved it to "2027-01", which selects AY2026-27.
  • Now it raises "cannot move the period '2026-01' by 1 year(s)". That stops the whole compile (uk_runtime/ledger_targets.py:508-516, :1453-1461) and the generators.
  • A month selector with an annual cutoff fails the same way.
• Cause and fix: my round-4 suggestion didn't distinguish match policies. The fix is to read the period the way the policy that consumes it does: typed for exact; for latest_not_after, typed where the type can read the spelling and untyped otherwise.
• The new local test (test_uk_ledger_targets.py:2812-2846) builds exactly this reference and calls it unmovable, so it pins the regression.
• Impact: latent, since no committed reference has both a typed selector and an observation period.
• Residuals:
  • the int fast path ignores the type (:4727-4728);
  • the constructor still validates the observation period untyped (:356-360), so a valid fiscal "2099-00" is refused;
  • with an untyped exact selector the match falls back to the fact's own type (:5092), so an untyped "2011-12" still restamps to "2012-12". The changelog's "as the exact-match path does" holds only for typed selectors.

2. The floor guard's case-sensitive constant pattern drops lowercase floors (packages/microcosm-build/tests/engine_free/uk/test_uk_receipt_counts.py:225).

• Now missed: rounds 3 and 4 caught lowercase names at the start of a line. This head misses:

  • floor = 10, support_floor=10, and small_cell_floor = 10;
  • sdc_floor = 10 and disclosure_floor: int = 10;
  • row_threshold = 10.

  No live instance is affected.

• Fix: the two false positives the reply credits to case sensitivity go away because of the new lookbehind (:222) and the narrowed key pattern (:218). So IGNORECASE can stay, with the words matched as snake-case tokens, e.g. (?:^|_)(?:min|minimum|floor|sdc|small|threshold)(?:_|\b).

• Docstring claims it doesn't meet:

  • #: comment continuations. One live, harmless wrap sits at uk_runtime/target_weights.py:101-102.
  • Float-, Optional- and Final-annotated defaults.
  • Single-quoted dict keys.

• False positives: fill-character format specs such as {n:0>9} are flagged, though they fail loudly.

• Brittleness: the kept phrases embed indentation and line breaks, so reflowing those comments fails the guard, again loudly.

3. Small: the observation-period tests require each period to be the January after the reference period, and call int() on that period. A fiscal or prefixed observation period, or a string target period like "2025-26", would fail them. That's fine for D5, but it deserves a comment in the test.

Checked and found true

• Typed shifts: every spelling each type can read moves correctly, with its start and range end both shifted. Untyped results are identical to round 4.
• Generators: both use the default --period of 2025 and the base of 2025. The DfE rows at 2024, 2025, 2026 and 2030 give "2025-01", "2026-01", "2027-01" and "2031-01".
• Graph: imports are clean, and the graph's loud failures on a missing boundary are intact.
• Guard: it scans the tree in 1.3 s with no catastrophic backtracking.
• Cross-PR: neither SPI donor savings rebased to SPI's 2023-24 bands (savings follow-up to #1418) #1428 nor UK: company cars from the FRS job table (dashboard #316) #1430 declares observation periods, and the guard flags none of their added lines.
• CI on 0da33d0cd: every job passes.

Disclosure-control items went to the author separately.

🤖 Generated with Claude Code

[juaristi22]

Thanks for the fifth review, of 0da33d0cd. Everything below is at 4af602d81.

New in 0da33d0cd

• 1. Restamps under the default policy: fixed. An observation period is now read as its match policy reads it: under the selector's period_type for exact, and untyped for the latest_not_after cutoff.
  • A fiscal- or academic-year selector's "2026-01" moves to "2027-01" again, and a month selector's "2026" to "2027".
  • The suggested rule for the cutoff, typed wherever the type can read the spelling, still picks wrong for one spelling. Take "2011-12" under the default policy:
    • the cutoff reads it as December 2011, and it reads the facts' own spellings untyped too;
    • under a fiscal-year selector it has to become "2012-13" to keep selecting the year's own fact;
    • under a calendar-year selector it has to become "2012-12".
  • So where the selector's type reads a spelling differently from the cutoff, the restamp now refuses. An untyped exact match refuses "2011-12" the same way, since it reads the period under each fact's own type. A prefixed or four-digit spelling reads one way.
  • The local compile test now uses "FY26". The parser reads it as the year 26, so it can't move under any policy.
  • Residuals:
    • the int path checks the year under the type, so a month type refuses a bare year;
    • the constructor checks an exact period under the selector's type, so a fiscal "2099-00" stands, and an exact fiscal "2026-01" is refused when the reference is built;
    • the changelog and docstrings no longer say "as the exact-match path does". They describe each policy's reading and the two refusals.
• 2. Floor guard: IGNORECASE is back, and the floor words (min, minimum, floor, sdc, small, threshold) are matched as snake-case words of a name.
  • All six missed names match again, and admin_count=10 and minutes=10 still don't.
  • Also caught now:
    • defaults under any annotation without a comma outside its brackets (float, Optional[int], Final);
    • single-quoted keys;
    • #: continuations. That found the wrap in uk_runtime/target_weights.py, which counts homes rather than records, so it's kept.
  • Format specs with a fill character ({n:0>9}) are skipped.
  • Kept phrases now match across any whitespace between their words, so re-indenting no longer fails them. Moving a line break across a comment marker still fails, loudly.
  • The new guard flags none of the lines that UK #1095 A/B fixes: the SPI age band (uk-data#628), households' HRAD from Table 8a receipts, and the SPI pay floor (uk-data#629) #1426, SPI donor savings rebased to SPI's 2023-24 bands (savings follow-up to #1418) #1428 or UK: company cars from the FRS job table (dashboard #316) #1430 add.
• 3. Observation-period tests: a comment now says D5's periods are untyped Januaries after an integer year, and that a fiscal, prefixed or range-spelt period needs its own expectation.

Parity-reason replay: my round-4 reply overstated it.

• The classifier's four reasons now live in one constant, PARITY_DIFFERENCE_REASONS in gates.py, with unchanged texts.
• The test starts each row from its kind's reason, then replays the tool's rules, so a changed classifier text and a removed rule fail too.
  • Removing the PIP rule fails the national receipt.
  • Changing the fixture-only text fails the national receipt too. The local receipt's generic rows are all calibration drift, so a change to that text fails both.

The touched test files pass locally, and each fix fails its tests under mutation. The gate digests, the coverage manifest check and the H2 fixture check pass too. CI runs the full tiers.